See every layer your AI runs on.
Then add to it.
AI and HPC operations
Orbit reads the datacenter, the compute, the fabric, the agents and the model providers, and renders one operating picture with a cost per completed task you can defend. When you need more compute, it attaches straight into the cluster you already run.
Read only against everything it observes. Deploys inside your boundary. No telemetry leaves your environment.
See every layer your AI runs on.
Then add to it.
AI and HPC operations
Orbit reads the datacenter, the compute, the fabric, the agents and the model providers, and renders one operating picture with a cost per completed task you can defend. When you need more compute, it attaches straight into the cluster you already run.
Read only against everything it observes. Deploys inside your boundary. No telemetry leaves your environment.
The Console
Multi-Tenancy, Single Pane of Glass
Operations, board builder, capacity and an attached workspace, reconstructed from the reference build.
Layers read, from rack power to model provider tokens
Canonical model, so boards survive a change of vendor
Ways to attach capacity: partition, node pool, workspace, endpoint
Write credentials held against any system Orbit observes
The Gap
Multi-Tenancy, Single Pane of Glass
Nobody is short of data. The questions that matter simply cross every boundary your tooling was built inside.
- Attribution
Who used that GPU hour?
Consumption cannot be assigned to the workflow or the team that caused it. Chargeback is estimated rather than measured, and capacity decisions are made without knowing what the capacity served.
- Cause
Why did that task fail?
It could be a provider rate limit, a routing decision, a saturated node, or a thermal event two racks away. Four tools, unsynchronised clocks, and most of the resolution time spent assembling context rather than fixing anything.
- Capacity
Do we need more, or better?
Without a unit cost you cannot tell an under-provisioned cluster from an inefficient workload. Both look like a queue. One is solved with procurement and the other is made worse by it.
One · See
Orbit reads every altitude and joins them to the unit of work.
Not a dashboard over your metrics. A model beneath them.
- ALT 05
Model Providers
Not a dashboard over your metrics. A model beneath them.
- ALT 04
Agents
Tasks, tool calls, retries, escalations
- ALT 03
Fabric
Routing, connector health, queue depth
One Operating Picture
On One Clock
One unit
of workOne Operating Picture
On One Clock
- ALT 02
Compute and Cloud
Accelerators, jobs, allocation, cloud
- ALT 01
Datacenters
Rack power, thermal, cooling
Five altitudes, one correlation identity, one picture. The vertical line is the part nobody else has.
When Something Breaks
A causal path, not a starting point for one.
Most of the time spent resolving an incident is not spent fixing anything. It is spent gathering context and correlating logs by hand across tools that do not share a clock. Orbit does that join when the reading arrives, so the answer is already assembled when the alert fires.
- Conditions raised on objective burn rate, not raw thresholds
- Every alert either carries an action or is marked informational
- One timeline across all five altitudes on a single normalised clock
Built to Outlive Your Infrastructure
Change vendors. Keep your boards.
Panels bind to meaning rather than to metric names. Swap accelerators, add a region, move a workload from your own racks to a managed endpoint, and the board keeps working. You change the sensor, not the board.
- Reads Supermicro, Dell and HPE platform managers where you already run them
- Reads out of band over Redfish, so it keeps reporting when a host goes down
- Coverage is reported honestly, never interpolated
- Every figure carries the confidence of the weakest join behind it
Unchanged. Not rebuilt, not migrated.
The Number Everything Else Rests On
Cost per completed task, and why it changes the next two sections.
Two consumption costs, at opposite ends of your stack, in incompatible units. Orbit is the only place they meet. Once they do, a capacity decision stops being a guess.
Cost per completed agent task, traced to the token and the GPU second.
Because every altitude resolves into one model, Orbit follows a single unit of work all the way down and prices it. Finance gets a defensible figure. Engineering gets the chain behind it. Both get the assumptions stated in the open rather than buried.
Orbit labels every component measured, attributed or apportioned. A figure is only as defensible as the join behind it, and we would rather tell you that than let you find out in review.
Two · Extend
Out of capacity on a Tuesday? Add some.
Attach an external accelerator pool from Cylix AI Cloud without leaving the console. Choose a class, choose a window, secure it. Billed by the hour from one hour to seven days.
A100
$1.05/hr
L40S
$1.95/hr
H100
$2.10/hr
H200
$2.40/hr
B200
$2.75/hr
Adding capacity makes your cost figure more accurate, not less. Rented accelerators are metered by the provider, so their cost is measured rather than estimated. Owned capacity is amortised and attributed. Run a meaningful share of work on metered capacity and a larger portion of your unit cost becomes a measurement rather than an attribution. Orbit shows you that shift instead of blending the two.
The Workspace
Root on a GPU box, one click from the console.
No ticket, no provisioning request, no waiting for a platform team to schedule you. Pick an image, pick how much external access it gets, and you are at a prompt.
- Egress defaults to an allowlist. A research container with open egress is a route out for your own data
- Sessions are recorded, because a root shell on rented capacity is the first thing an auditor asks about
- The same SSH details work from your own terminal
Editions
Run it yourself, or let us run it.
Two editions, sold differently because they carry different operating responsibility.
On Premise
Your infrastructure, your boundary
Deployed on your own hardware or cloud account by our DevOps team. One operator, no tenancy, nothing shared with anyone. Operate it yourself or have us operate it for you under the managed model.
Cylix organizations whose regulatory posture requires physical separation, and anyone who simply prefers to hold the keys.
Hosted Managed Private Cloud
We run the control plane
Cylix operates the service and you get a tenant within it. Each tenant runs in its own isolated namespace with its own quota, network policy and service account. Reserved accelerators are contractual and are never shared.
The control plane is shared with logical separation, and we say so plainly rather than implying more isolation than exists. If that is not acceptable to you, take the on-premise edition instead.
For Compute Service Providers
Offer an operations panel alongside your capacity.
If you sell accelerated compute, Orbit is the layer that makes it more than a commodity.
- Create
Tenants in minutes
Name, plan, regions. Provisioning creates an isolated namespace with quota and network policy, then verifies the isolation holds before the tenant is usable.
- Operate
See across, or step inside
Watch capacity, utilisation and margin across every tenant, or enter a tenant context to support them. Provider functions are unavailable inside a tenant context, and every action records both identities.
- Differentiate
Compute with a defensible unit cost
Each customer sees their own allocation, consumption and spend, and nothing of their neighbours. Capacity at a given specification is close to a commodity. Capacity with an operating picture is not.
The things people ask
on the first call.
Not to anything it observes, and it never will. Every sensor holds read-scoped credentials and the software has no write path to an observed system. Securing capacity and attaching it are writes to our own service on your instruction, which is a different thing, and we keep that boundary explicit rather than blurring it.
No. Your monitoring platform watches machines and services. Orbit watches the join between altitudes and prices the unit of work that crosses them. Keep your platform for what it already does well and point Orbit at the correlation problem it was never built to solve.
No. Orbit reads whatever request lifecycle and model routing you already run. The fabric is the reference implementation because we built both, and the join is tightest there, but Orbit reads a generic OpenAI-compatible endpoint just as well.
No. Orbit deploys inside your boundary and reads locally. What leaves is a correlation identity and a set of numbers, never the underlying request, response or document content.
As accurate as the join behind it, and we label which parts of it are measured, attributed or apportioned so you can see the difference rather than take a single blended number on faith.
Yes, for Slurm and Kubernetes. Rented nodes federate into the cluster or controller you already run, so users submit work the same way they did before Orbit existed.
Logical isolation at the namespace, quota, network policy and service account level, on a shared control plane. If that boundary is not sufficient for your regulatory posture, the on-premise edition gives you physical separation instead.
A working session against your own environment, not a slide deck. We map where a fabric would apply first, model the cost difference against what you run today, and leave you with a staged adoption path, including the parts where the honest answer is not yet.
Connect With Us
Take the next step in your AI journey. Reach out to our sales team to discuss your upcoming AI project or connect with our support team for assistance.
